Network Effects in AI
Tags:
Market Landscape and Context
In present day market environments, network effects drive about 70% of value for technology companies and technology-enabled services. Emerging trends in AI point to many potential areas where network effects can amplify product utility and customer adoption, and in turn drive a large portion of market value for technology companies. It can be hugely valuable in early product development to architect a product to allow users to participate in value creation, and effectively let the use of the product add value to the other users. Every incremental customer can effectively add value to each existing customer, as well as each future customer.
The traditional model of network effects contends that the network of customers utilizing a product provides the majority of the value of the project, versus the product itself. Prime examples are in the social media space, with industry titans like Facebook and Instagram, where their existing networks of users deliver the most value for each incremental user that signs up. Companies like Craigslist and Ebay illustrate network effects in its purest form, as they rely entirely on their network of users to deliver their value proposition, as evidenced by the success of their marketplaces in congruence with their bare interfaces (all of their value is derived from the network of users interacting on their tech-enabled marketplace platforms, not in branding, marketing, or any other product dimension).
New Types of Network Effects in AI
Furthermore, a recent survey from Gartner shows that 37% of organizations worldwide have adopted some form of AI, which is a growth rate of 270% within the past four years. It appears that, with the emergence of data sharing within networks and novel AI applications, new types of network effects may come about and bring value to AI startups specifically. For example, new types of network effects come about with new network participants that interact with & benefit from a startup’s AI, which brings about new categorizations for the network effect itself.
Take for example an AI startup that creates a robotic arm to replace restaurant workers at fast-casual restaurants. The restaurant “enterprises” themselves form a network of direct customers of the AI startup, adopting the robotic arm technology for operational enhancements. Simultaneously, the customers in the restaurant (the restaurants themselves serving as the first layer) are also part of the network, and also benefit from the cleanliness and improved preparation of the food (assuming the startup’s product works and their value-prop is substantiated) among other positive externalities (e.g. improved health outcomes). This secondary layer (with delivery platforms, other restaurant components, serving as tertiary layers) essentially benefits from the positive externalities of the AI’s introduction into the business.
The interaction of these layers appears to introduce a new type of network effect particular to AI startups (and tied to the business impacts of automation), which could be described as vertical network effects (similar in concept to the “vertical integration” of a business’s operations), across the chain of consumption.
An example that demonstrates the vertical nature of AI network effects is a startup that provides robotic farming automation for picking fruits and vegetables. A business that vertically integrates is aiming to operationally control the production of vegetables from farm, to wholesale distribution, to retail, to customer. Network effects play a role in a similar vertical fashion here, where the farm’s harvesting operations are automated and enhanced, downstream distributors benefit from greater quantities and higher quality produce with peak ripeness and lower wholesale purchase costs, and further upstream, fertilizer companies are benefitting from the operational expansion of the farm’s capabilities, which require larger purchases of fertilizer. The network of buyers of the produce at the end of the consumption chain also benefit from the improved produce quality and enhanced ripeness. Thus, across vertical network participants, network effects arise from a stimulus of transactions, consumption, and economic activity.
To complete this particular lens of viewing network effects, the conventional model of network effects would indicate more of a horizontal network effect, among the “same type” of network participants, i.e. within the category of fertilizer types (in farming AI), delivery types (in restaurant AI), and individual users (in social media).
New Categories of Network Effects in AI
With emerging AI technologies that are creating entirely new markets and redefining existing methodologies, there appear to also be new categories of network effects that come with it. One distinct category that has emerged is adaptive learning in AI, with startups building technologies that adaptively learn with new data in searching for an optimum, whether that be an optimal decision or optimal product. This could be a farming automation startup that is building an AI to detect peak ripeness, for example. Another category appears to be network effects that purely arise from existing networks of customers or users. The social media platforms described above, e.g. Facebook, Twitter, Snapchat, would fit this model. It’ll be interesting to see how new categories of network effects are defined with each new AI that goes-to-market.
Implications for Business Operations & Go-To-Market Strategies
Recognition of the immense positive impacts and value of network effects in AI should impact how a business operates, goes-to-market, and scales. With network effects in mind, business strategy should be developed with the direct consumers of the product in mind, as well as the indirect beneficiaries of the product, which can occur downstream or upstream (e.g. in the farming automation example), which in turn create positive feedback loops that reinforce business operations and sales channels at each stage.
Network Effects in Data Generation
Network effects in general add value to a business with each new node or user that is added. A data network effect specifically occurs when newly added data continuously improves a product or service. As the usage of the product grows, the data contributed by each user grows, and the product improves with the aggregated data. Data network effects provide an inherent advantage to a vast array of companies, especially those that are AI/ML based, by allowing them to leverage their customer data more effectively. Data network effects can be derived from the data generated from large, anonymous groups of people that engage with the company’s products. Recommendation systems, for example, benefit from the network effects stemming from the large and continuously growing supply of data that is generated from the user base. The deeper the data set, the better the recommendations that can be made.
To provide further granularity, data generation also provides positive network effects for each “participant category.” One participant category is the set of users of the product, which largely generate the data in question. Another auxiliary category is the set of secondary beneficiaries or those positively affected by the product. The larger dataset improves the product, which in turn results in an enhanced product experience for the initial set of users. This positive feedback loop creates a data network effect via data generation or aggregation.
Data network effects can be described as a flywheel feedback loop, with customers generating data that in turn informs prediction models & algos, which in turn brings in more customers. Data continues to aggregate and the process amplifies. Moreover, businesses can unlock immense competitive advantages by utilizing AI/ML, given the potential for competitive moats with data network effects. Data network effects can build defensibility and add value through increased user retention and alignment to user preferences through prediction models and algorithms.
Network Effects in Synthetic Data
Data network effects can also play a role in synthetic data generation. Synthetic data products adaptively learn as more data is generated, and the convenience of availability is a huge value proposition. Apple, Google, and Microsoft have all published research that illustrates how synthetic data is highly convenient for machine learning and training models.
Moreover, Synthetic data appears to help startups compete with large-cap corporations that tend to have swsthes of data readily available. The ability to generate synthetic data creates a level playing field for startups to compete with data-rich giants.
The adoption of synthetic data in AI practice is growing, and alongside that trend is a massive market opportunity for synthetic data platforms that can generate convenient and readily available data. Google researchers are now training AI in simulated worlds for example.
(Addendum)
Other Use Cases
Data network effects can also play a role for platforms that are building autonomous vehicles. Tesla for example could benefit from strong data network effects, as its fleet of cars gathers data that drives continuous improvement of Tesla’s AI/ML, which could then be applied to navigation, automation, and other decision making processes.
Roadmap for Leveraging Data Network Effects & Potential Limitations
AI/ML products tend to inherently exhibit data network effects, but not all companies are able to leverage these effects to their advantage. Complexity of data poses one limitation, leading to the growing importance of data management and data scientists. Further downstream, the potential for competitive advantage also is dependent on the specific use case. While recommender systems are a popular use case where AI/ML has provided companies with data network effects that give rise to competitive advantages, the same positive effects are hard to unlock in facial recognition, home devices, and certain subareas of robotics.